Layered architecture
Schema, core, communication, infrastructure, and observability layers stay cleanly separated.
Distributed & Trustworthy AI
Trustworthy AI at Scale
FedPilot is a Ray-backed platform for topology-aware federated learning and Trustworthy AI research. Define schemas, virtual nodes, and adaptation policies before the runtime materializes actors and placement groups.
Distributed AI
Train globally, keep data local
Federated learning coordinates model updates across clients without centralizing raw data — the foundation of privacy-preserving distributed AI.
Cross-silo & cross-device
Institutions to edge fleets
Non-IID & heterogeneity
FedProx, sampling, robust agg
Vision & NLP models
CNN, ResNet, BERT, custom PyTorch
Privacy budgets
DP-SGD, secure aggregation
Platform design
Four core contributions
FedPilot treats distributed systems concerns as first-class — not hidden behind FL-only abstractions.
Systems abstractions
Lazy virtual-node materialization, topology-aware routing, and ICRF as a core primitive.
Topology adaptation
Data-driven clustering from label distributions drives placement and horizontal scaling.
Grounded observability
OpenTelemetry, Prometheus, Grafana, and Streamlit capture pressure and network I/O as experiment artifacts.
Distributed systems
Inter-Cluster Ray Fabric (ICRF)
The ICRF is the spine of multi-cluster federation: one logical graph, hybrid transport chosen automatically per hop.
Ray shared memory · intra-cluster
HTTP / Ray Serve · inter-cluster
Clustering wires the fabric; HybridAdjacencyMatrix encodes routes; HybridTopologyManager enforces them at runtime.
Documentation map
Explore by operational layer
The docs mirror how FedPilot runs in production — from boot configuration through telemetry.
01
Entry & Configuration
Getting started, CLI usage, and the configuration reference.
Open layer → 02Orchestration & Infrastructure
Ray virtual nodes, topology manager, and global object store.
Open layer → 03Schemas & Applications
Schemas SDK and AppFactory for mapping paradigms to engines.
Open layer → 04Federated Core & Communication
ICRF, FederatedBase, aggregators, compression, Shapley analysis.
Open layer → 05Tool Registries
Plugin registries for models, topology, metrics, and adaptation.
Open layer → 06Security & Privacy
Differential privacy and cryptography / secure aggregation.
Open layer → 07Dashboards & Telemetry
Metrics export, Ray and Streamlit dashboards, deployment guide.
Open layer →Ready to run an experiment?
Install FedPilot, configure a topology, and ship reproducible FL runs on a laptop or Ray cluster.